RFR: JDK-8166648,,jib make run-test for langtools results in intermittent failures on windows-x86

[Erik Joelsson]

On 2016-10-10 14:56, David Holmes wrote:
> On 10/10/2016 6:53 PM, Erik Joelsson wrote:
>> Hello,
>>
>> The variable ARCH_DATA_MODEL is not initialized anywhere. I know JPRT
>> sets it, but I don't think we should rely on that.
>
> ./autoconf/buildjdk-spec.gmk.in:ARCH_DATA_MODEL := 
> $(OPENJDK_BUILD_CPU_BITS)
>
A test makefile does not necessarily run in the context of a configure 
configuration (and he buildjdk-spec will never be in effect when running 
a test anyway). In this case, the check for 32 bit has to happen in the 
makefile.

/Erik

> David
>
>> /Erik
>>
>>
>> On 2016-10-08 03:34, Tim Bell wrote:
>>> Hi Jon
>>>
>>> On 10/07/16 15:48, Jonathan Gibbons wrote:
>>>> Tim,
>>>>
>>>> The problem only appears on 32-bit Windows, but the fix appears to be
>>>> across the board.
>>>> Does this penalize or otherwise adversely affect test runs on other
>>>> platforms?
>>>
>>> Good point.  Here is an updated webrev where serialGC is requested
>>> only on 32 bit platforms:
>>>
>>>    http://cr.openjdk.java.net/~tbell/8166648/webrev.01/
>>>
>>> This change tested OK in JPRT on the following platforms:
>>>   solaris_sparcv9_5.11-product-c2-langtools_jtreg
>>>   solaris_x64_5.11-product-c2-langtools_jtreg
>>>   linux_i586_3.8-product-c2-langtools_jtreg
>>>   linux_x64_3.8-product-c2-langtools_jtreg
>>>   macosx_x64_10.9-product-c2-langtools_jtreg
>>>   windows_i586_6.3-product-c2-langtools_jtreg
>>>   windows_x64_6.3-product-c2-langtools_jtreg
>>>
>>> I inspected the test logs to verify -XX:+UseSerialGC is passed only
>>> during 32-bit testing.
>>>
>>> Tim
>>>
>>>>
>>>> -- Jon
>>>>
>>>> On 10/07/2016 07:15 AM, Tim Bell wrote:
>>>>> Again with the review link .. see below:
>>>>>
>>>>> On 10/07/16 06:40, Tim Bell wrote:
>>>>>> Hello
>>>>>>
>>>>>> The change of HotSpot runtimes on 32-bit Windows [1] has been a 
>>>>>> source
>>>>>> of langtools test failures due to memory space exhaustion.
>>>>>>
>>>>>> On our Windows build/test clients, the combination of the C2
>>>>>> runtime and
>>>>>> parallel GC causes about 20 extra threads to be started in each VM.
>>>>>> These added thread stacks, plus heap management in C2, result in too
>>>>>> much memory pressure to fit the workload into 32-bit VMs.
>>>>>>
>>>>>> This fix will run langtools jtreg tests using the serial garbage
>>>>>> collector instead of using the default collector (parallel GC).
>>>>>
>>>>> http://cr.openjdk.java.net/~tbell/8166648/
>>>>>
>>>>>>
>>>>>> Testing:
>>>>>>    -5 out of 5 jobs (with the fix) passed langtools testing
>>>>>>    -4 out of 5 control jobs (without the fix) failed testing in
>>>>>> windows_i586_6.3-product-c2-langtools_jtreg
>>>>>>
>>>>>>
>>>>>> Thanks in advance-
>>>>>>
>>>>>> Tim
>>>>>>
>>>>>> [1] JDK-8154209 "Remove client VM from default JIB profile on
>>>>>> windows-x86 and linux-x86"
>>>>>> https://bugs.openjdk.java.net/browse/JDK-8154209
>>>>>
>>>>
>>>
>>